Knowledge work
Coding agents
Agentic coding with a named harness: Terminal-Bench, DeepSWE, and SWE-Atlas-QnA as reported by Artificial Analysis.
As of 13 August 2026. Conditional pack. Artificial Analysis Coding Agent Index, 9 Jul 2026. GPT-5.6 Sol, Terra, and Luna in Codex. Harness is part of the treatment.
The Coding Agent Index pairs models with agentic harnesses on DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA. Codex versus Claude Code is a different system. Index points are not official SWE-bench Verified.
AA reports GPT-5.6 Sol (max) in Codex at 80 index points, Terra 77, Luna 75. Sol leads the three underlying evals in that write-up, tying Grok 4.5 on SWE-Atlas-QnA. Per-eval percentages are not copied where AA did not publish them as numbers.
- Scale
- Artificial Analysis Coding Agent Index, observed Jul 2026
- Provenance
- Artificial Analysis Coding Agent Index, 9 Jul 2026. GPT-5.6 Sol, Terra, and Luna in Codex. Not an Antidote SWE-bench Verified run.
- Contamination
- High. Public coding benches. Harness choice (Codex vs Claude Code) changes the cell.
- Cannot establish
- A software-vendor product score, or substitution of an engineering org.
Metrics, kept separate
- Coding Agent Index
- Terminal-Bench
- DeepSWE
- SWE-Atlas-QnA
- Cost per task
Third-party leaderboard · not an Antidote run · not a product score
Artificial Analysis, 9 Jul 2026. GPT-5.6 family in Codex. Index points shown as a fraction of 100. Not official SWE-bench Verified. Fable 5 index points are not copied; AA states Sol leads this index versus Fable 5 in Claude Code on cost.
| System | Coding Agent Index |
|---|---|
| GPT-5.6 Sol max / Codex | 80% |
| GPT-5.6 Terra max / Codex | 77% |
| GPT-5.6 Luna max / Codex | 75% |
What the figures bound
On this dated AA ladder, the GPT-5.6 family is the current public coding-agent observation that includes Sol, Terra, and Luna. Cost per task is reported as lower for Sol in Codex than Fable 5 in Claude Code, at roughly 40% cheaper.
That is not a software-vendor diligence score and not a reason to skip sealed engineering workflows.
What this page does not conclude
- Official SWE-bench Verified for these models
- A software-product or services-org replacement rate
- Fable 5 index points, which AA did not publish as a number in the cited article
- An Antidote coding-agent run
Cited as calibration on these memos — not a score of those products.
Not attached to a software, advisory, or automation memo yet.